Under the Hood · Part 4 · 4 min read
The wheels
A model without tools is a brain that can only talk. Tools are where the talking turns into doing.
By Aikansh ·
Imagine the most brilliant advisor in the world, locked in a room with no phone, no computer, no door. Ask anything and the answer is dazzling. But they cannot send the email, book the flight, run the numbers, or change a single thing in the actual world. They can only tell you what they'd do.
That is a language model on its own. Text in, text out — and not one thing more. Last time we told it who to be. But a driver, however sharp, still isn't going anywhere without wheels. This piece is about the wheels: the part that finally lets it touch the road.
The thing that took me too long to see
The leap from a chatbot to an agent isn't a smarter model. It's tools — and a tool is a plainer idea than the word suggests. It's just a door to the outside world: one thing the model is allowed to actually do instead of merely describe. Run this code, query this database, read this file, send this message. Give it none, and the genius stays locked in the room, narrating. Give it a few, and the talking turns into doing.
HOW A TOOL ACTUALLY WORKS
- The model can only produce text. Left alone it will write you the shell command, the SQL, the email — perfectly — and then sit there. Writing it and running it are two different things.
- A tool is a door to the world. Each one is a single capability — run_tests, read_file, send_email, search_web. The model now has hands, not just a mouth.
- It doesn't press the button — it asks. It emits a structured request ("call run_tests"); something outside it does the running and hands the result back. The model reads that result and keeps going.
- MCP is a standard socket, not the wheels. Build a tool once and any model that speaks MCP can use it — a drawer of proprietary cables replaced by one standard plug. Spreading fast; not required; not the mechanism.
Once you see it
Every time you've watched an AI do something real — edit your files, pull live data, actually send the thing — you weren't watching a smarter model. You were watching a model handed wheels. And every time one felt useless, trapped, all-talk-no-action, it usually wasn't dumb. It just had nothing to reach the world with.
It's the whole reason the assistant I run day to day — Donna — is useful at all: she can read my calendar and my task list and act on them. Take those tools away and she's a very polite chatbot, narrating what she would do if only she could reach something.
A model without tools is
a brain that can only talk.
TRY IT — 3 MINUTES
Open Claude Code (or any agent that can run tools). Ask it two things that reach different doors to the world:
- "read the files in this folder and tell me what this project does" — the file-system door.
- "look up today's headlines and summarize them" — the open-web door.
Watch the exact moment it stops talking and calls a tool — opening a real file, reaching a real website. That call, punching from text into the world, is the whole difference between a chatbot and an agent.
It can act now. The next question is the one that decides whether the acting is any good: how does it know if what it just did actually worked? That needs a way to keep score — which is next.
FOR THE CURIOUS — WHAT'S ACTUALLY HAPPENING
Mechanically: the model is given a set of tool definitions — name, description, expected inputs — alongside your prompt. When it decides to use one, it doesn't run anything; it returns a structured tool-call request. Something executes it and appends the result back into the context as a new observation, which the model reads on its next step.1 For most tools that "something" is the harness you're running (client tools); a handful of built-ins run on the provider's own infrastructure instead (server tools, e.g. web search) — a real distinction, because it changes who holds the keys.
MCP — the Model Context Protocol, published by Anthropic in November 2024 — standardizes how those tools and data sources are exposed to a model, so an integration built once works across clients.2 Adoption has been fast, but it is one way to wire tools up, not a precondition for agentic action.
The honest hard part: tools are where the real-world risk moves in. A model that can only talk can, at worst, be wrong. One that can delete a file, send a message, or spend money can be wrong and costly. Permissions, approvals, and tight scoping aren't red tape — they're the line between an assistant and a liability. The wheels make the car useful; they're also what let it crash.
SOURCES & NOTES
- Anthropic, tool use (function calling) documentation — the model emits a structured tool-call request; the caller executes it and returns the result as a new observation; client vs. server tools.
- Anthropic, "Introducing the Model Context Protocol" (Nov 25, 2024) — an open standard for connecting models to tools and data; one way to wire tools, not required.